mirror of
https://github.com/mattjaybe/SillyTavern-Pathweaver.git
synced 2026-04-26 10:30:47 +00:00
v1.5.0 - Added Reasoning Mode
- New setting: Reasoning Mode. Increases max tokens and properly handles thinking tags. - Reasoning Mode allows for setting the max output token
This commit is contained in:
mattjaybe
GitHub
parent
00078c8ca1
commit
5e23475b4c
3 changed files with 159 additions and 13 deletions
143
index.js
143
index.js
|
|
@ -128,7 +128,9 @@
|
|||
surprise_depth_min: 2, // minimum messages away (used for random range or fixed min)
|
||||
surprise_depth_max: 6, // maximum messages away (used for random range or fixed max)
|
||||
surprise_randomize: true, // randomize depth between min and max
|
||||
surprise_endless: false // auto-rearm a new surprise after each one fires
|
||||
surprise_endless: false, // auto-rearm a new surprise after each one fires
|
||||
reasoning_mode: false, // Enable reasoning mode for models like DeepSeek
|
||||
max_output_tokens: 16384 // Configurable max output tokens (default 16K for reasoning models)
|
||||
});
|
||||
|
||||
|
||||
|
|
@ -587,14 +589,32 @@ GUIDELINES:
|
|||
let userPrompt = '';
|
||||
let calculatedMaxTokens = 0;
|
||||
|
||||
// Calculate base tokens needed for suggestions
|
||||
const tokensPerSuggestion = settings.suggestion_length === 'long' ? 400 : 280;
|
||||
const baseTokensNeeded = settings.suggestions_count * tokensPerSuggestion + 800;
|
||||
|
||||
if (settings.reasoning_mode) {
|
||||
// Reasoning mode: use configurable max_output_tokens, but ensure it's at least enough for suggestions
|
||||
// Reasoning models need more tokens for their thinking process
|
||||
calculatedMaxTokens = Math.max(settings.max_output_tokens || 8192, baseTokensNeeded);
|
||||
log(`Reasoning mode enabled. Using max_tokens: ${calculatedMaxTokens} (config: ${settings.max_output_tokens})`);
|
||||
} else {
|
||||
// Normal mode: use the original formula with reasonable bounds
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, baseTokensNeeded));
|
||||
}
|
||||
|
||||
if (category === 'director' && customDirections?.length) {
|
||||
if (mode === 'story_beats') {
|
||||
// Story Beats: 1 input = 1 suggestion (Classic behavior)
|
||||
const dirList = customDirections.map((d, i) => `${i + 1}. ${d}`).join('\n');
|
||||
userPrompt = `[STORY CONTEXT]\n${contextBlock}\n\n[TASK]\nGenerate exactly ${customDirections.length} suggestions, one for each of the following directions.\n\nUSER DIRECTIONS:\n${dirList}\n\nFORMAT:\n[EMOJI] TITLE\nDESCRIPTION\n\nGUIDELINES:\n- PREVENT BLEED: Each suggestion must be strictly isolated to its corresponding input beat. Do NOT combine events from different beats unless explicitly requested.\n- Follow the specific direction for each suggestion EXACTLY.\n- Keep titles punchy and plain text (no asterisks).\n- ${settings.suggestion_length === 'long' ? 'Write 4-6 sentences per suggestion.' : 'Write 2-3 sentences per suggestion.'}\n- Do NOT include any preamble.${settings.stream_suggestions ? '\n\nSTREAMING: Output one complete suggestion at a time. Each suggestion MUST start with [EMOJI] TITLE then DESCRIPTION; end each with --- before the next. Do NOT repeat a title or copy content from one suggestion into another. Every suggestion is independent and self-contained.' : ''}`;
|
||||
// Be generous with token budgeting; otherwise the model may stop early (e.g. 4/6 suggestions).
|
||||
const tokensPerSuggestion = settings.suggestion_length === 'long' ? 400 : 280;
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, customDirections.length * tokensPerSuggestion + 800));
|
||||
// Recalculate for director mode with custom directions
|
||||
const dirTokensNeeded = customDirections.length * tokensPerSuggestion + 800;
|
||||
if (settings.reasoning_mode) {
|
||||
calculatedMaxTokens = Math.max(settings.max_output_tokens || 8192, dirTokensNeeded);
|
||||
} else {
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, dirTokensNeeded));
|
||||
}
|
||||
} else {
|
||||
// Single Scene: Combined inputs = N suggestions (New behavior)
|
||||
const combinedDirections = customDirections.join(' ');
|
||||
|
|
@ -603,9 +623,12 @@ GUIDELINES:
|
|||
: 'Each description should be 2-3 sentences, concise but evocative.';
|
||||
|
||||
userPrompt = `[STORY CONTEXT]\n${contextBlock}\n\n[TASK]\nThe user has provided the following direction/scenario for the next scene:\n"${combinedDirections}"\n\nBased on this direction, generate exactly ${settings.suggestions_count} DISTINCT options or variations for how this scene could play out.\n${lengthInstruction}\n\nFORMAT:\n[EMOJI] TITLE\nDESCRIPTION\n\nGUIDELINES:\n- All suggestions must follow the user's direction but offer different execution/flavor.\n- Keep titles punchy and plain text.\n- Do NOT include any preamble.${settings.stream_suggestions ? '\n\nSTREAMING: Output one complete suggestion at a time. Each suggestion MUST start with [EMOJI] TITLE then DESCRIPTION; end each with --- before the next. Do NOT repeat a title or copy content from one suggestion into another. Every suggestion is independent and self-contained.' : ''}`;
|
||||
// Be generous with token budgeting; otherwise the model may stop early (e.g. 4/6 suggestions).
|
||||
const tokensPerSuggestion = settings.suggestion_length === 'long' ? 400 : 280;
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, settings.suggestions_count * tokensPerSuggestion + 800));
|
||||
// Recalculate for director single scene mode
|
||||
if (settings.reasoning_mode) {
|
||||
calculatedMaxTokens = Math.max(settings.max_output_tokens || 8192, baseTokensNeeded);
|
||||
} else {
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, baseTokensNeeded));
|
||||
}
|
||||
}
|
||||
} else {
|
||||
const lengthInstruction = settings.suggestion_length === 'long'
|
||||
|
|
@ -613,8 +636,12 @@ GUIDELINES:
|
|||
: 'Each description should be 2-3 sentences, concise but evocative.';
|
||||
|
||||
userPrompt = `[STORY CONTEXT]\n${contextBlock}\n\n[TASK]\nGenerate exactly ${settings.suggestions_count} distinct suggestions.\n${lengthInstruction}\nFollow the format specified in the system instructions exactly.\nIMPORTANT: Use PLAIN TEXT for titles - do NOT wrap titles in **asterisks**.\nDo NOT include any preamble.${settings.stream_suggestions ? '\n\nSTREAMING: Output one complete suggestion at a time. Each suggestion MUST start with [EMOJI] TITLE then DESCRIPTION; end each with --- before the next. Do NOT repeat a title or copy content from one suggestion into another. Every suggestion is independent and self-contained.' : ''}`;
|
||||
const tokensPerSuggestion = settings.suggestion_length === 'long' ? 400 : 280;
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, settings.suggestions_count * tokensPerSuggestion + 800));
|
||||
// Use the pre-calculated baseTokensNeeded (already calculated above)
|
||||
if (settings.reasoning_mode) {
|
||||
calculatedMaxTokens = Math.max(settings.max_output_tokens || 8192, baseTokensNeeded);
|
||||
} else {
|
||||
calculatedMaxTokens = Math.min(8192, Math.max(2048, baseTokensNeeded));
|
||||
}
|
||||
}
|
||||
|
||||
let result = '';
|
||||
|
|
@ -817,7 +844,11 @@ GUIDELINES:
|
|||
if (!text) return [];
|
||||
|
||||
// First, strip any reasoning/thinking tags from the entire response
|
||||
// Handles both XML-style (DeepSeek, etc.) and HTML-style tags
|
||||
let cleanedText = text
|
||||
// XML-style reasoning tags: <think>, </think>, <thinking>, </thinking>, etc.
|
||||
.replace(/<\/?(thought|think|thinking|reasoning|reason)\s*[^>]*>/gi, '')
|
||||
// HTML-style reasoning tags
|
||||
.replace(/<(thought|think|thinking|reasoning|reason)>[\s\S]*?<\/\1>/gi, '')
|
||||
.replace(/<(thought|think|thinking|reasoning|reason)\/>/gi, '')
|
||||
.replace(/<(thought|think|thinking|reasoning|reason)\s*\/>/gi, '')
|
||||
|
|
@ -828,8 +859,8 @@ GUIDELINES:
|
|||
const suggestions = [];
|
||||
let blocks = [];
|
||||
|
||||
// Broad emoji pattern that catches most emojis
|
||||
const emojiRegex = /[\u{1F300}-\u{1F9FF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F600}-\u{1F64F}\u{1F680}-\u{1F6FF}\u{2300}-\u{23FF}\u{2B50}\u{1FA00}-\u{1FAFF}]/gu;
|
||||
// Broad emoji pattern that catches most emojis including extended ranges
|
||||
const emojiRegex = /[\u{1F000}-\u{1F02B}\u{1F0A0}-\u{1F0FF}\u{1F100}-\u{1F1FF}\u{1F300}-\u{1F9FF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F600}-\u{1F64F}\u{1F680}-\u{1F6FF}\u{2300}-\u{23FF}\u{2B00}-\u{2BFF}\u{2B50}\u{1FA00}-\u{1FAFF}]/gu;
|
||||
|
||||
// Strategy 1: Split by --- separator (various formats)
|
||||
blocks = cleanedText.split(/\n---\n|\n---|---\n|\n\n---\n\n/);
|
||||
|
|
@ -871,6 +902,9 @@ GUIDELINES:
|
|||
|
||||
// Strip any remaining reasoning tags from this block
|
||||
trimmed = trimmed
|
||||
// XML-style reasoning tags
|
||||
.replace(/<\/?(thought|think|thinking|reasoning|reason)\s*[^>]*>/gi, '')
|
||||
// HTML-style reasoning tags
|
||||
.replace(/<(thought|think|thinking|reasoning|reason)>[\s\S]*?<\/\1>/gi, '')
|
||||
.replace(/<[^>]*>/g, '')
|
||||
.trim();
|
||||
|
|
@ -937,12 +971,15 @@ GUIDELINES:
|
|||
// STREAMING: incremental parse and UI
|
||||
// ============================================================
|
||||
|
||||
const emojiRegexStream = /[\u{1F300}-\u{1F9FF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F600}-\u{1F64F}\u{1F680}-\u{1F6FF}\u{2300}-\u{23FF}\u{2B50}\u{1FA00}-\u{1FAFF}]/gu;
|
||||
const emojiRegexStream = /[\u{1F000}-\u{1F02B}\u{1F0A0}-\u{1F0FF}\u{1F100}-\u{1F1FF}\u{1F300}-\u{1F9FF}\u{2600}-\u{26FF}\u{2700}-\u{27BF}\u{1F600}-\u{1F64F}\u{1F680}-\u{1F6FF}\u{2300}-\u{23FF}\u{2B00}-\u{2BFF}\u{2B50}\u{1FA00}-\u{1FAFF}]/gu;
|
||||
|
||||
/** Parse a single suggestion block into { emoji, title, description } or null */
|
||||
function parseOneBlock(blockText) {
|
||||
if (!blockText || typeof blockText !== 'string') return null;
|
||||
let trimmed = blockText
|
||||
// XML-style reasoning tags
|
||||
.replace(/<\/?(thought|think|thinking|reasoning|reason)\s*[^>]*>/gi, '')
|
||||
// HTML-style reasoning tags
|
||||
.replace(/<(thought|think|thinking|reasoning|reason)>[\s\S]*?<\/\1>/gi, '')
|
||||
.replace(/<[^>]*>/g, '')
|
||||
.trim();
|
||||
|
|
@ -2429,6 +2466,26 @@ GUIDELINES:
|
|||
<p class="pw_setting_hint pw_setting_stream_hint">
|
||||
Cards appear as each suggestion is generated. Works with Ollama and OpenAI-compatible APIs; Connection Profile may also support streaming.
|
||||
</p>
|
||||
|
||||
<!-- Reasoning Mode Settings -->
|
||||
<div class="pw_setting_row" style="margin-top: 12px;">
|
||||
<span class="pw_setting_label"><i class="fa-solid fa-brain"></i> Reasoning Mode</span>
|
||||
<div class="pw_toggle ${settings.reasoning_mode ? 'active' : ''}" data-setting="reasoning_mode"></div>
|
||||
</div>
|
||||
<p class="pw_setting_hint">
|
||||
Enable for reasoning models (DeepSeek, etc.). Increases max tokens and properly handles thinking tags.
|
||||
</p>
|
||||
|
||||
<div id="pw_modal_max_tokens_row" class="pw_setting_row" style="${settings.reasoning_mode ? 'display: flex' : 'display: none'}; margin-top: 8px;">
|
||||
<span class="pw_setting_label"><i class="fa-solid fa-terminal"></i> Max Output Tokens</span>
|
||||
<div class="pw_setting_control">
|
||||
<input id="pw_modal_max_output_tokens" type="number" class="text_pole" value="${settings.max_output_tokens || 16384}" min="512" max="128000" step="512" style="width: 100px;"
|
||||
title="Maximum tokens for reasoning model output">
|
||||
</div>
|
||||
</div>
|
||||
<p id="pw_modal_max_tokens_hint" class="pw_setting_hint" style="${settings.reasoning_mode ? 'display: block' : 'display: none'}; margin-top: 4px;">
|
||||
Higher values allow more thinking tokens for reasoning models. 8K-32K recommended.
|
||||
</p>
|
||||
</div>
|
||||
|
||||
<div class="pw_settings_section">
|
||||
|
|
@ -2689,6 +2746,19 @@ GUIDELINES:
|
|||
createActionBar();
|
||||
}
|
||||
|
||||
// Reasoning mode: show/hide max tokens input
|
||||
if (setting === 'reasoning_mode') {
|
||||
if (settings.reasoning_mode) {
|
||||
jQuery('#pw_modal_max_tokens_row').show();
|
||||
jQuery('#pw_modal_max_tokens_hint').show();
|
||||
} else {
|
||||
jQuery('#pw_modal_max_tokens_row').hide();
|
||||
jQuery('#pw_modal_max_tokens_hint').hide();
|
||||
}
|
||||
// Also sync to extension panel
|
||||
jQuery('#pw_max_output_tokens_row').toggle(settings.reasoning_mode);
|
||||
}
|
||||
|
||||
saveSettings();
|
||||
syncSettingsToPanel(); // Sync to extension panel (NOW after logic)
|
||||
});
|
||||
|
|
@ -2725,6 +2795,15 @@ GUIDELINES:
|
|||
// Suggestion length
|
||||
jQuery('#pw_sm_suggestion_length').on('change', function () { settings.suggestion_length = this.value; saveSettings(); syncSettingsToPanel(); });
|
||||
|
||||
// Modal: Max output tokens
|
||||
jQuery('#pw_modal_max_output_tokens').on('change', function () {
|
||||
const val = parseInt(this.value) || 16384;
|
||||
settings.max_output_tokens = Math.max(512, Math.min(128000, val));
|
||||
this.value = settings.max_output_tokens;
|
||||
saveSettings();
|
||||
syncSettingsToPanel();
|
||||
});
|
||||
|
||||
// Surprise Me: depth min select
|
||||
jQuery('#pw_sm_surprise_depth_min').on('change', function () {
|
||||
const val = parseInt(this.value) || 2;
|
||||
|
|
@ -3953,6 +4032,14 @@ GUIDELINES:
|
|||
jQuery('#pw_include_description').prop('checked', settings.include_description);
|
||||
jQuery('#pw_include_worldinfo').prop('checked', settings.include_worldinfo);
|
||||
jQuery('#pw_stream_suggestions').prop('checked', settings.stream_suggestions);
|
||||
// Reasoning mode settings
|
||||
jQuery('#pw_reasoning_mode').prop('checked', settings.reasoning_mode);
|
||||
jQuery('#pw_max_output_tokens').val(settings.max_output_tokens || 16384);
|
||||
if (settings.reasoning_mode) {
|
||||
jQuery('#pw_max_output_tokens_row').show();
|
||||
} else {
|
||||
jQuery('#pw_max_output_tokens_row').hide();
|
||||
}
|
||||
// Surprise Me
|
||||
jQuery('#pw_surprise_randomize').prop('checked', settings.surprise_randomize);
|
||||
jQuery('#pw_surprise_endless').prop('checked', settings.surprise_endless);
|
||||
|
|
@ -4017,6 +4104,16 @@ GUIDELINES:
|
|||
jQuery('.pw_toggle[data-setting="include_scenario"]').toggleClass('active', settings.include_scenario);
|
||||
jQuery('.pw_toggle[data-setting="include_description"]').toggleClass('active', settings.include_description);
|
||||
jQuery('.pw_toggle[data-setting="include_worldinfo"]').toggleClass('active', settings.include_worldinfo);
|
||||
|
||||
// Reasoning mode settings
|
||||
jQuery('.pw_toggle[data-setting="reasoning_mode"]').toggleClass('active', settings.reasoning_mode);
|
||||
jQuery('#pw_max_output_tokens').val(settings.max_output_tokens || 16384);
|
||||
if (settings.reasoning_mode) {
|
||||
jQuery('#pw_max_output_tokens_row').show();
|
||||
} else {
|
||||
jQuery('#pw_max_output_tokens_row').hide();
|
||||
}
|
||||
|
||||
// Update provider visibility
|
||||
jQuery('#pw_sm_profile_box, #pw_sm_ollama_box, #pw_sm_openai_box').hide();
|
||||
if (settings.source === 'profile') jQuery('#pw_sm_profile_box').show();
|
||||
|
|
@ -4217,6 +4314,28 @@ GUIDELINES:
|
|||
syncSettingsToModal();
|
||||
});
|
||||
|
||||
// Reasoning mode toggle
|
||||
jQuery('#pw_reasoning_mode').on('change', function () {
|
||||
settings.reasoning_mode = this.checked;
|
||||
saveSettings();
|
||||
syncSettingsToModal();
|
||||
// Show/hide max_output_tokens setting based on reasoning mode
|
||||
if (this.checked) {
|
||||
jQuery('#pw_max_output_tokens_row').show();
|
||||
} else {
|
||||
jQuery('#pw_max_output_tokens_row').hide();
|
||||
}
|
||||
});
|
||||
|
||||
// Max output tokens
|
||||
jQuery('#pw_max_output_tokens').on('change', function () {
|
||||
const val = parseInt(this.value) || 16384;
|
||||
settings.max_output_tokens = Math.max(512, Math.min(128000, val));
|
||||
this.value = settings.max_output_tokens;
|
||||
saveSettings();
|
||||
syncSettingsToModal();
|
||||
});
|
||||
|
||||
// Include Scenario
|
||||
jQuery('#pw_include_scenario').on('change', function () {
|
||||
settings.include_scenario = this.checked;
|
||||
|
|
|
|||
|
|
@ -6,7 +6,7 @@
|
|||
"js": "index.js",
|
||||
"css": "style.css",
|
||||
"author": "mattjaybe",
|
||||
"version": "1.4.6",
|
||||
"version": "1.5.0",
|
||||
"homePage": "https://github.com/mattjaybe/SillyTavern-Pathweaver",
|
||||
"auto_update": true
|
||||
}
|
||||
|
|
|
|||
|
|
@ -183,6 +183,33 @@
|
|||
Connection Profile may also support streaming.
|
||||
</p>
|
||||
|
||||
<!-- Reasoning Mode Settings -->
|
||||
<label class="checkbox_label pw_checkbox_row pw_stream_row" for="pw_reasoning_mode"
|
||||
style="margin-top: 12px;">
|
||||
<input id="pw_reasoning_mode" type="checkbox" class="checkbox"
|
||||
title="Enable for reasoning models like DeepSeek">
|
||||
<span><i class="fa-solid fa-brain"></i> Reasoning Mode</span>
|
||||
</label>
|
||||
<p class="pw_setting_hint" style="margin: 4px 0 0 0; font-size: 0.8rem; opacity: 0.85;">
|
||||
Enable for reasoning models (DeepSeek, etc.). Increases max tokens and properly handles
|
||||
thinking tags.
|
||||
</p>
|
||||
|
||||
<div id="pw_max_output_tokens_row" class="pw_row_2col" style="margin-top: 10px; display: none;">
|
||||
<div>
|
||||
<label for="pw_max_output_tokens" class="pw_input_label">
|
||||
<i class="fa-solid fa-terminal"></i> Max Output Tokens
|
||||
</label>
|
||||
<input id="pw_max_output_tokens" type="number" class="text_pole" value="16384" min="512"
|
||||
max="128000" step="512" style="width: 100%;"
|
||||
title="Maximum tokens for reasoning model output">
|
||||
</div>
|
||||
<div></div>
|
||||
</div>
|
||||
<p id="pw_max_output_tokens_hint" class="pw_setting_hint" style="margin: 4px 0 0 0; font-size: 0.75rem; opacity: 0.7; display: none;">
|
||||
Higher values allow more thinking tokens for reasoning models. 8K-32K recommended.
|
||||
</p>
|
||||
|
||||
<h4 class="pw_section_title" style="margin-top: 15px; font-size: 0.9rem;">
|
||||
<i class="fa-solid fa-book-open"></i> Context Sources
|
||||
</h4>
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue